Inverse synthetic aperture radar(ISAR),as the core technology of high-resolution radar imaging,has significant application value in fields of marine monitoring,air reconnaissance and space target recognition due to it...Inverse synthetic aperture radar(ISAR),as the core technology of high-resolution radar imaging,has significant application value in fields of marine monitoring,air reconnaissance and space target recognition due to its all-weather and long-range advantages.However,the current research on ISAR target classification mainly focuses on the low-frequency band,and the target classification methods for terahertz ISAR images is still in its infancy.Compared with low-frequency band ISAR images,terahertz ISAR images not only have richer pixel-level details,but also have more complex scattering characteristics,which poses new challenges to the feature extraction and classification capabilities of existing methods.To address these challenges,this paper proposes a terahertz ISAR image target classification method based on deep learning and designs a multi-scale space-frequency dual-branch features fusion(MSFF)network.The network consists of two key blocks:the multi-scale spatial feature extraction(MSFE)block and the frequency domain convolution awareness(FDCA)block.The MSFE block extracts local spatial features of different scales through multi-scale convolution kernels,thereby enhancing the model's perception ability of the local features of target.The FDCA block extracts the global features of the target through frequency domain transformation,enhancing the model's ability to capture global structural information.This method achieves feature fusion between spatial and frequency-domain branches through feature concatenation,effectively integrating the multi-domain information and significantly improving the performance of feature extraction.Furthermore,the MSFF network provides fewer parameters and floating-point operations per second(FLOPs).Experimental results on the selfconstructed terahertz ISAR dataset demonstrate that the MSFF network achieves a minimum classification accuracy of 99.49%,with only 0.49 M parameters and 0.37 G FLOPs.Furthermore,comparative experiments with the low-frequency ISAR datasets further validate the superior performance of the MSFF network.展开更多
基金supported in part by the National Natural Science Foundation of China under Grant No.62588201in part by the Artificial Intelligence Promotes Scientific Research Paradigm Reform and Empowers Discipline Advancement Planin part by the Natural Science Foundation of Shanghai under Grant No.21ZR1444300。
文摘Inverse synthetic aperture radar(ISAR),as the core technology of high-resolution radar imaging,has significant application value in fields of marine monitoring,air reconnaissance and space target recognition due to its all-weather and long-range advantages.However,the current research on ISAR target classification mainly focuses on the low-frequency band,and the target classification methods for terahertz ISAR images is still in its infancy.Compared with low-frequency band ISAR images,terahertz ISAR images not only have richer pixel-level details,but also have more complex scattering characteristics,which poses new challenges to the feature extraction and classification capabilities of existing methods.To address these challenges,this paper proposes a terahertz ISAR image target classification method based on deep learning and designs a multi-scale space-frequency dual-branch features fusion(MSFF)network.The network consists of two key blocks:the multi-scale spatial feature extraction(MSFE)block and the frequency domain convolution awareness(FDCA)block.The MSFE block extracts local spatial features of different scales through multi-scale convolution kernels,thereby enhancing the model's perception ability of the local features of target.The FDCA block extracts the global features of the target through frequency domain transformation,enhancing the model's ability to capture global structural information.This method achieves feature fusion between spatial and frequency-domain branches through feature concatenation,effectively integrating the multi-domain information and significantly improving the performance of feature extraction.Furthermore,the MSFF network provides fewer parameters and floating-point operations per second(FLOPs).Experimental results on the selfconstructed terahertz ISAR dataset demonstrate that the MSFF network achieves a minimum classification accuracy of 99.49%,with only 0.49 M parameters and 0.37 G FLOPs.Furthermore,comparative experiments with the low-frequency ISAR datasets further validate the superior performance of the MSFF network.
